Description
Privacy and communication cost concerns have led to interest in federated learning (FL) for edge machine learning applications. While standard machine learning has a rich ecosystem of distributed training libraries, federated learning’s novelty means researchers lack the necessary frameworks to efficiently explore the large design space unique to FL. In this work we propose, build, and evaluate a benchmarking framework for FL algorithms. Our system, RayLEAF, allows users to train FL algorithms in parallel while testing model compression approaches in a simulated federated setting.